The Anthropic Claude Code team publicly shares their internal skills usage experience, summarized into nine major categories.

Gate News, March 18 — Thariq, a member of the Anthropic programming tool Claude Code team, shared their internal experience using Skills. Skills is the most commonly used extension mechanism for Claude Code, with hundreds currently active within Anthropic. The team categorizes them into nine main types: 1. Libraries and API references: teaching agents to correctly use internal libraries, CLI, and SDKs, with code snippets and common pitfalls; 2. Product validation: verifying code output with tools like Playwright, tmux, etc., recording videos as evidence, and inserting assertions at each step; 3. Data collection and analysis: connecting monitoring systems, built-in credentials, dashboard IDs, and common query templates; 4. Business process automation: automating repetitive tasks such as stand-up summaries, ticket creation, and weekly reports with a single command; 5. Code scaffolding: generating project templates according to team standards, suitable for scenarios where natural language requirements cannot be fully covered by code alone; 6. Code quality and review: running automatically as hooks or GitHub Actions, including derived sub-agents for adversarial review; 7. CI/CD and deployment: monitoring PRs, retrying CI, progressive releases, and automatic rollbacks; 8. Runbook: troubleshooting through multiple tools based on alerts or error signatures, producing structured reports; 9. Infrastructure operations: safe operations like orphan resource cleanup, dependency approval, and cost investigation. The team emphasizes key practices including: Skills are not just markdown files but complete folders that can contain scripts, assets, and data, leveraging the file system for incremental information disclosure; the most valuable content is the Gotchas section, which should continuously accumulate agent failure modes; the description field in Skills is meant as a trigger condition for the model, not a human-readable summary; historical logs can be stored within Skills directories to enable cross-session memory. The team tracks Skill usage data through the PreToolUse hook to measure trigger rates and effectiveness.